christian pellegrin wrote: > On Mon, Jan 18, 2010 at 8:40 AM, Brian Murphy <[email protected]> wrote: >>> >> I can back this up. We have decided not to use mcp2515 because of packet >> loss >> on our board. >> > > Hi, > > out of curiosity what architecture have you worked on? With the > mcp2515 there are good chances that you can keep up with a full stream > (because of the 10 mhz SPI interface and, most important, the ability > to send an entire Rxed CAN packet in a row). > > Anyway I fully support Wolfgang's claims about performance. > Unfortunately the SPI infrastructure is rather critical and on some > architectures it's just plain bad. For example on S3C24XX setting DMA > up takes ages and the shift register is 1 byte deep for IRQ mode, so > the best option you have is just wait every single transfer to finish > with an overall decrease in processor throughput. Another problem is > that the best SPI drivers aren't always in the kernel mainline. > > As far as the mcp251x driver is concerned, now that it is in the > mainline, I'm looking forward to substitute worqueues with threaded > interrupts (which is something we all have to get used to sooner or > later) and adding hardware level CAN-address filtering. Both of these > could make it more useful.
I also see space for improvements. As the mcp252x does just buffers two received messages (one shadow buffer, IIRC) it's important to read out the messages as quick as possible. Wolfgang. _______________________________________________ Socketcan-core mailing list [email protected] https://lists.berlios.de/mailman/listinfo/socketcan-core
